[Paper Review] Identification of Light Sources using Machine Learning

Chenglong You, Mario A. Quiroz‐Juárez+7 more|arXiv (Cornell University)|Sep 17, 2019

Spectroscopy Techniques in Biomedical and Chemical Research1 references4 citations

TL;DR

This paper proposes a machine learning approach using a single artificial neuron (ADALINE) and naive Bayes classifiers to identify light sources—specifically coherent and thermal light—at the single-photon level. By leveraging statistical fluctuations in photon counts, the method reduces required measurements by several orders of magnitude, achieving accurate discrimination with fewer than ten measurements at mean photon numbers below one.

ABSTRACT

The identification of light sources represents a task of utmost importance for the development of multiple photonic technologies. Over the last decades, the identification of light sources as diverse as sunlight, laser radiation and molecule fluorescence has relied on the collection of photon statistics or the implementation of quantum state tomography. In general, this task requires an extensive number of measurements to unveil the characteristic statistical fluctuations and correlation properties of light, particularly in the low-photon flux regime. In this article, we exploit the self-learning features of artificial neural networks and naive Bayes classifier to dramatically reduce the number of measurements required to discriminate thermal light from coherent light at the single-photon level. We demonstrate robust light identification with tens of measurements at mean photon numbers below one. Our work demonstrates an improvement in terms of the number of measurements of several orders of magnitude with respect to conventional schemes for characterization of light sources. Our work has important implications for multiple photonic technologies such as LIDAR and microscopy.

Motivation & Objective

• Address the challenge of identifying light sources with minimal measurements, especially in low-photon flux regimes.
• Overcome the limitations of conventional methods like quantum state tomography and Mandel parameter estimation, which require extensive data.
• Demonstrate that machine learning can drastically reduce the number of measurements needed for light source discrimination.
• Explore the feasibility of using simple models like ADALINE and naive Bayes for real-time, low-resource photonic applications.
• Enable practical implementations in LIDAR, microscopy, and remote sensing by minimizing data acquisition needs.

Proposed method

• Train a single ADALINE neuron using photon count statistics from coherent and thermal light sources.
• Use the surjective photon counting method to record TTL pulses from superconducting nanowire single-photon detectors (SNSPDs).
• Apply the ADALINE model, an adaptive linear neuron, to classify light sources based on statistical fluctuations in photon arrival times.
• Compare performance with a naive Bayes classifier trained on the same photon statistics.
• Evaluate additional models including a one-dimensional convolutional neural network (1D-CNN) and a multilayer neural network (MNN) for benchmarking.
• Use cross-entropy loss with scaled conjugate gradient backpropagation for training, and employ ReLU and softmax activation functions in deep networks.

Experimental results

Research questions

• RQ1 Can machine learning reduce the number of measurements required to identify coherent versus thermal light sources?
• RQ2 How does the performance of a simple ADALINE neuron compare to more complex models like 1D-CNNs and MNNs in low-photon regimes?
• RQ3 What is the minimum number of measurements needed for reliable light source identification at mean photon numbers below one?
• RQ4 Can the ADALINE model effectively distinguish signal photons from ambient noise with minimal data?
• RQ5 How does model complexity affect generalization and computational cost in photon statistics classification?

Key findings

• The ADALINE model achieved accurate light source identification with fewer than ten measurements at mean photon numbers below one, representing a reduction of several orders of magnitude compared to conventional methods.
• The ADALINE model outperformed the naive Bayes classifier in terms of recognition accuracy under the same experimental conditions.
• The 1D-CNN and MNN models showed comparable performance to ADALINE but required significantly more computational resources and were more complex to train.
• For mean photon numbers of 0.4 and 0.77, the MNN’s accuracy decreased with increasing hidden layer neurons due to over-parameterization and poor generalization.
• The 1D-CNN achieved near-saturated accuracy faster than the MNN as data size increased, indicating faster convergence in low-photon regimes.
• Overall accuracy improved with more training data points, but the 1D-CNN reached peak performance more rapidly than the MNN, despite higher computational cost.